Interpretable machine learning model stratifies latent tuberculosis risk

Researchers in China have developed a machine learning model that can tell the difference between active tuberculosis in the lungs and latent tuberculosis infection using only information collected through questionnaires. The model can also divide people with latent infection into three risk groups to help health workers decide who needs closer follow-up. The study involved several medical centers and was published in the open-access journal BMC Public Health on 22 October 2026. It included 782 people from 21 designated tuberculosis institutions across Henan Province. Of these, 401 had latent tuberculosis infection, while 381 had newly diagnosed pulmonary tuberculosis and had not yet received treatment.

About one-quarter of the world's population is estimated to have latent tuberculosis infection. This means the bacteria are present in the body, but the person does not have active disease. Current screening tests can miss some cases, and many existing prediction models depend on laboratory tests that cost money. To address these problems, the researchers chose simple, low-cost factors that health workers could collect easily, even in basic health clinics. They gathered information through standardized, face-to-face questionnaires and randomly divided the data into a training group and a testing group in a 7:3 ratio. Using a statistical method called LASSO regression with ten-fold cross-validation, they selected 11 factors for the model. They then built and compared seven machine learning algorithms. To measure how well these models worked, they used several tests, including the area under the receiver operating characteristic curve (AUC), F1 score, Brier score, calibration curves, and decision curve analysis.

The eXtreme Gradient Boosting model, called XGB, performed best on the test data. Its AUC was 0.818 (95% confidence interval [CI] 0.763–0.873), its F1 score was 0.764 (95% CI 0.698–0.822), and its Brier score was 0.176 (95% CI 0.148–0.204). To make the results useful in clinics, the researchers created three risk levels using a method based on percentiles. A predicted probability above 0.763 meant high risk, while a probability below 0.328 meant low risk. People whose scores fell between these cutoffs belonged to the middle-risk group. The share of people who actually had pulmonary tuberculosis increased sharply from one group to the next. In the low-risk group, 20.00% had pulmonary tuberculosis, compared with 84.29% in the high-risk group.

This large difference between the groups is the study's main practical finding. Instead of treating everyone with latent tuberculosis infection as having the same level of risk, the model identifies a smaller group in which active tuberculosis is much more common. This could help health workers decide who should receive further testing and who might need preventive care. The researchers stress that the model is meant to help prioritize people for further assessment, not to diagnose tuberculosis on its own. They also designed the model to make its predictions easier to understand.

The team used a method called SHapley Additive exPlanations, or SHAP, to show how each factor affected the model's predictions. The analysis found that the most influential factors were being male, having a low body mass index (BMI), being older, losing weight unintentionally during the previous six months, and having a low annual household income. The researchers also made the model available as a free online calculator. This could make it easier to use in primary care clinics and public health services that have limited laboratory resources. However, the authors clearly explain that the study has important limitations.

The researchers used a case-control study, comparing people with newly diagnosed, untreated pulmonary tuberculosis with people who had latent infection. Therefore, the model identifies differences between people who already have active disease and those who have latent infection. It does not predict which people with latent infection will develop active tuberculosis in the future. The online calculator is described as a research prototype. Before it can be used to guide decisions about preventive treatment, it must be tested in other groups of people and evaluated in studies that follow participants over time. The training and testing data also came from just one province, so it is not yet known how well the model will work in other populations, places with different levels of tuberculosis transmission, or different health systems.

Despite these limitations, the model's ability to distinguish between the two groups, its good agreement between predicted and observed risks, its use of SHAP to explain predictions, and its reliance on low-cost information make it a possible tool for grouping people with latent tuberculosis infection by risk. Because the study is openly available, other research teams can also try to repeat the findings using their own data.

Subject of Research: Developing and testing an understandable machine learning model to group people with latent tuberculosis infection by risk.

Article Title: Development and validation of an interpretable machine learning model for risk stratification among individuals with latent tuberculosis infection: a multicenter case-control study

Article References: Wang, W., Sun, D., Zhang, Y., Xu, X., Chang, W., Zheng, D., He, M., Xie, C., Qi, Y., Xu, Y., Zhu, J., & Diao, L. (2026). Development and validation of an interpretable machine learning model for risk stratification among individuals with latent tuberculosis infection: a multicenter case-control study. BMC Public Health. https://doi.org/10.1186/s12889-026-30066-9

Image Credits: AI generated.

DOI: 10.1186/s12889-026-30066-9

Keywords: latent tuberculosis infection, pulmonary tuberculosis, machine learning, XGBoost, SHAP, risk grouping, LASSO regression, case-control study, epidemiological factors, web calculator, BMC Public Health, development.